Klaviyo Agent Guidance: Guardrails Before You Unleash Autonomous Agents
Klaviyo’s 2026 release wave is bigger than the usual feature drop. The company repositioned its AI work as an autonomous B2C CRM and shipped more than 75 new features across marketing, data, and analytics. Two agents anchor that story. Composer is an agentic interface where a marketer describes a goal in plain language, say reactivate lapsed customers, and it builds a launch-ready campaign: audience segmentation, messaging, and channel orchestration across email and SMS. The other is K:AI Customer Agent, which handles service and sales work like order tracking, returns and exchanges, subscription management, and loyalty lookups.
The catch is obvious once you sit with it. Something that drafts your campaigns and answers your customers in real time is also something that can damage your brand if the tone is off or it fails to escalate when it should. That is what Agent Guidance is for. It is the control layer where you define tone, escalation rules, and when human intervention is required, so the sensitive and complex interactions get handled the right way. This guide covers how to set those guardrails up, and the point is to do it before you turn the agents loose, not after the first bad transcript lands in your inbox.
Start With Tone, Not Capability
Most teams open Agent Guidance and immediately start toggling what the agent is allowed to do. Wrong order. Configure tone first, because Customer Agent may handle thousands of conversations a day and every one of them is your brand talking. The good news is you write tone rules in plain language, no code required.
Take a DTC skincare brand. You might write: keep replies warm and first-person, lead with empathy before policy when a customer reports irritation, and never promise a clinical outcome or use words like cure or guaranteed. Those three lines alone reshape how the agent sounds. A premium watch or leather-goods brand would invert that: stay measured and precise, skip exclamation points and slang, and stay calm and confident on authentication questions rather than defensive.
Write tone rules as specifically as you can. Vague instructions like be friendly and professional give the agent nothing to act on. Hand it a banned-words list, the exact greeting you want on a complaint, and the form of address to use. The more concrete the rule, the more reliably the agent follows it. Treat the tone block as your brand voice guide compressed into instructions an agent can actually execute.
Escalation Rules: When the Agent Should Step Back
Tone alone is not enough. The expensive failures come from an agent that does not know when to stop. Escalation rules in Agent Guidance handle that. You define which scenarios the agent is not allowed to resolve on its own and must hand up.
Here is a table of common DTC scenarios you can configure against directly:
| Scenario | Agent Guidance rule | Agent behavior |
|---|---|---|
| Routine order or shipping status | Full automation allowed | Looks up and answers, no human touched |
| Standard return within policy window | Automate, but log anything over 500 USD | Creates the return, issues a label |
| Customer is angry or uses words like lawsuit or chargeback | Escalate immediately, no auto-reply | Stops, routes to a human, attaches the thread |
| Compensation or refund beyond policy | Escalate to a supervisor | Makes no promise, says a specialist will follow up |
| Allergy, injury, or product-safety claim | Force human takeover | Holds at once, flags it, notifies compliance |
The table is there to pull out the scenarios that get expensive the moment the agent gets them wrong, not to automate everything you can. Allergy and injury claims, for instance, should never be agent-handled, full stop. Klaviyo lets you trigger on keywords, dollar thresholds, and sentiment signals, so your job is to decide which lines your brand cannot afford to cross.
One thing teams miss: escalation is not the same as dumping the problem. A well-built rule has the agent package the conversation context and hand it over, so the human who picks it up does not have to re-interview the customer. When you write the rule, include carry the full transcript as part of the handoff.
Make the Human-Handoff Conditions Specific
Escalation rules decide whether to hand off. Human-handoff conditions decide how, and to whom. Keep those two separate in your head when you configure them.
In practice I set three tiers. Tier one is a soft prompt: the agent is unsure but nowhere near a red line, so it adds a line like let me confirm that with a colleague, notifies the team asynchronously, and does not block the customer. Tier two is a hard transfer: any red-line scenario from the table above drops straight into the human queue and the agent goes silent. Tier three is the fallback: if a customer asks for a human two or three turns running, transfer regardless of topic. Forcing a chat-bound customer to keep arguing with an agent is the fastest way to burn goodwill.
Build in a time dimension too. You can set the agent to handle inquiries after hours but, for anything needing a person, tell the customer a human will reply by 9 am and open a ticket. That keeps a customer from sitting at midnight with an agent that just keeps deflecting.
Composer needs guardrails of the same kind even though it does not face customers directly. Cap the discount it can apply on its own, block reactivation campaigns from touching anyone who has already unsubscribed, and require human review on any copy that makes a price or delivery promise before it sends. Composer can take a campaign all the way to launch-ready on its own, but whether it actually sends, and whether a person looks first, is a gate you keep your hand on.
Pilot Before You Open the Floodgates
Once the guardrails are written, do not switch everything on. The advantage of this setup is that you can start narrow. Let Customer Agent handle only the safest bucket first, order and shipping status, run it for a week or two, then read the actual transcripts one by one. Watch for whether the tone holds and whether anything that should have escalated slipped through.
Check three things on each read: did it use a banned word, did the red-line scenarios route to a human as instructed, and did the handoff carry clean context. When a rule misfires, go back into Agent Guidance, fix it, then widen the scope. Once that loop runs smoothly, hand it the harder buckets like returns and loyalty.
Same discipline for Composer. Use it to build one or two low-risk campaigns first, like a new-product teaser to your active segment, review before sending, and confirm the segmentation and copy hold up before you let it near win-back or abandoned-cart flows.
Last thing: guardrails are not set-and-forget. Your policies change, you launch new products, you hit new complaint types, and your Agent Guidance rules have to keep up. Treat it as something you come back to on a schedule, not a form you fill out on launch day. The more autonomous the agents get, the more this layer is your last line of defense.
Related Articles
Attentive Thread 2026: What the Agentic AI and RCS Upgrades Mean Before BFCM
At Thread 2026 on May 26, Attentive revealed that nearly half its customers are already running AI agents. Before BFCM, Brand Voice 2.0, AI Email Template Generation, RCS Visibility AI, and AI Grow 2.0 are all shipping. Here is what each one does and how to prepare.
Intercom Fin for E-Commerce: How AI Turns Support Into a Sales Channel
Launched May 8, 2026, Fin for Ecommerce handles pre-purchase discovery and post-purchase support in a single conversation thread. Shopify integration takes minutes. Early numbers show 10% conversation-to-order conversion and 20% higher AOV on Fin-assisted orders.
Pacvue Agent: Governed AI Execution for Amazon Retail Media
Pacvue Agent, launched April 14 2026, moves Amazon retail media from analysis to recommendation to governed execution in one workflow. This guide covers how sellers set guardrails, run plain-language AMC queries, and decide what to delegate.